From flatland to 3D: How AI is unlocking new chemical space
Drug discovery is entering a new era where synthetic chemistry and AI are working hand in hand to help scientists explore molecular territory that was, until recently, almost off the map.
In our recent CAS Insights webinar, AI-driven advancements in medicinal chemistry, we walked through a real-world case study with Dr. Philipp Natho (University of Bari, Italy) and Dr. Lee Walsh-Benn (CAS) that shows how this convergence is already producing results. They discussed a faster, more economical path from a novel scaffold to a validated bioactive compound.
From knowledge recording to knowledge prediction
Dr. Natho opened with a brief history of how chemists interacted with scientific knowledge. The print era required library access and weeks of manual searching. Digitization in the 1990s removed the geographic barrier but left chemists searching through unstructured PDFs. Curated databases like CAS SciFinder® turned document retrieval into data retrieval, delivering answers in seconds.
We are now, Dr. Natho argued, at a second inflection point: “It’s really been a shift from data retrieval to data inference.”
In this AI/ML era, curated databases serve not only as repositories but as training data for models that predict reactivity, synthetic routes, and binding affinity. This changes the chemist’s job from making compounds and hoping for an application to designing them with a hypothesis already in hand.
Escaping “flatland”
The case study sits at the heart of one of medicinal chemistry’s biggest current shifts: the move from flat aromatic scaffolds toward three-dimensional structures that better match biological binding sites.
Spiro[3.3]heptanes have already been validated as bioisosteres for several saturated heterocycles. Their smaller cousins, spiro[2.3]hexanes, occupy similar chemical space but are dramatically underrepresented — with one analogue, 1,5-dioxaspiro[2.3]hexane, returning just 47 products across seven publications and zero patents in CAS SciFinder. Dr. Natho’s group faced a “chicken-and-egg question”: Are these compounds absent because they have no medicinal value, or because no one has had a good way to create them?
A three-step workflow combining chemistry and AI
With more than 60 compounds in hand, the team built a three-step workflow to find a biological application:
- Unsupervised learning clustered the new spiro[2.3]hexanes against ~80 popular medicinal chemistry heterocycles using DFT-derived descriptors.
- AI-driven target prediction in CAS BioFinder® prioritized which analogues to make.
- In vitro validation in SH-SY5Y neuroblastoma cells confirmed the predictions and compound 81 showed the highest binding activity.
Dr. Natho summarized the impact: “Some of these compounds, I’m convinced, would have never been designed by rational drug design in the first place.”
The workflow saved roughly $15,000 in external testing and compressed the timeline from project design to biological validation to just eight months.
CAS BioFinder: Beyond a single case study and into agentic AI
Dr. Lee Walsh-Benn followed with a wider review of how CAS BioFinder fits into the drug discovery process:
- Predictive screening for potency and safety
- Matched molecular pair analysis for lead optimization
- Exploration of nearby chemical space in 2D and 3D
- Integration with CAS SciFinder for retrosynthesis and route planning
A more recent addition to CAS BioFinder and CAS SciFinder is CAS Newton℠, an agentic AI assistant trained on the CAS Content Collection™. Unlike general-purpose AI agents, CAS Newton grounds every answer in peer-reviewed literature and is designed to return null results honestly when the data isn’t there.
Final thoughts
What this case study demonstrates is a different way of running a discovery project. Instead of making compounds and hoping for an application, or screening blindly against panels of receptors, chemists can now use AI to focus their synthetic and experimental efforts on the candidates most likely to succeed and to confidently explore chemical spaces that would otherwise have stayed unexplored.
As Dr. Walsh-Benn noted in the Q&A, predictive analytics is not a verdict; it’s an arrow. Used that way, AI sharpens a chemist’s judgment rather than replacing it.
Watch the full webinar
To hear directly from Dr. Natho and Dr. Walsh-Benn, see the synthetic schemes, CAS BioFinder screenshots, and full Q&A discussion, watch the recording of the webinar.
Questions from the webinar
Is the predicted pActivity in CAS BioFinder a standard metric or specific to BioFinder?
The pActivity score is designed to be closely related to other activity measurements, but it is specific to the predictive models built into CAS BioFinder. It should be interpreted as a relative ranking within the platform.
Were any of the compounds predicted as inactive actually tested?
Yes. One compound that did not return a positive prediction for the μ-opioid receptor was tested anyway and showed a reasonable EC₅₀ in vitro. “No predicted activity” is not the same as “inactive,” and it usually means the model lacks sufficient training data in that region of chemical space. Predictions point you in the right direction; they don’t rule things out.
Where does AI fit into the typical R&D workflow?
Across all of it. In CAS BioFinder, AI accelerates the parts of the workflow that consume the most time: filtering literature and patents, screening compounds against target panels, prioritizing candidates for synthesis, and exploring scaffolds similar to those that have failed in earlier trials. With agentic tools like CAS Newton, scientists can also ask complex, cross-disciplinary questions in natural language and receive cited, literature-grounded answers in minutes.
Is the AI prediction based on structure-based or ligand-based drug design?
Both. More detail on how the predictions and confidence scores are calculated is available in the supporting information of the Angewandte Chemie article.
